Evaluating Dynamic Knowledge Encoding: Experimental Setup for Multi-Hop Logical Reasoning

Table of Links

Abstract and 1. Introduction

1. Background

2. Method

3. Experiments

   4.1 Multi-hop Reasoning Performance

   4.2 Reasoning with Distractors

   4.3 Generalization to Real-World knowledge

   4.4 Run-time Analysis

   4.5 Memorizing Knowledge

4. Related Work

5. Conclusion, Acknowledgements, and References

\ A. Dataset

B. In-context Reasoning with Distractors

C. Implementation Details

D. Adaptive Learning Rate

E. Experiments with Large Language Models

4 Experiments

Setup We conduct our experiments on three datasets focusing on multi-hop logical reasoning over natural language knowledge: ProofWriter [73], which measures the model’s ability to emulate reasoning over facts and rules expressed in natural language; CLUTRR-SG [28], which is generated from the CLUTRR [71] benchmark, a logical reasoning task that involves reasoning over family relationships between entities grounded in first-order logical proofs; and FOLIO [29], a reasoning benchmark with first-order logical reasoning problems written by expert annotators based on real-world knowledge. Each problem in these datasets requires multiple reasoning hops to answer.[1]

\ We compare our method against the following baselines: (1) a fine-tuned model that performs a forward pass on only the question without access to the knowledge (No-Facts), (2) a fine-tuned model that performs a forward pass on only the knowledge without access to the question (No-Question), (3) a model trained using RECKONING with random knowledge that is not relevant to the questions (Random-Facts), and (4) an ICR baseline that concatenates the knowledge K with the question x in a single context and is trained using supervised learning to predict the answer (FT-ICR). Our first three baselines sanity-check whether any surface-level patterns in the questions and facts can be exploited to make accurate predictions. The last baseline compares RECKONING to the conventional way of reasoning with language models. Unless stated otherwise, we use the GPT-2-small [59] model (∼124M parameters) as our initialization and refer by RECKONING to our method trained with the multi-task objective. We compute each score from the average across three different runs. For more details on the implementation, datasets, and examples, see Appendix A and Appendix C.

\

:::info Authors:

(1) Zeming Chen, EPFL (zeming.chen@epfl.ch);

(2) Gail Weiss, EPFL (antoine.bosselut@epfl.ch);

(3) Eric Mitchell, Stanford University (eric.mitchell@cs.stanford.edu)';

(4) Asli Celikyilmaz, Meta AI Research (aslic@meta.com);

(5) Antoine Bosselut, EPFL (antoine.bosselut@epfl.ch).

:::

:::info This paper is available on arxiv under CC BY 4.0 DEED license.

:::

[1] In ProofWriter, the number of reasoning hops is called the proof depth. To unify the presentation of the results, we use the term “hop” to describe the number of reasoning steps for both datasets.